AI literacy is not a training problem

Organisations reach for training as the default response to a capability gap. For most new software, this works: a workshop, a quick-reference card, and a week of practice cover most of what the user needs. AI is different. The gap is not procedural — it is a mental-model gap — and training programmes built on the procedural assumption consistently fail to close it.

What the user actually needs to know

Using AI well requires a set of judgements that have no analogue in traditional software. What does this tool actually do, as distinct from what it appears to do? When is its output reliable and when is it confidently wrong (see AI’s most dangerous failure mode is confident wrongness)? What context does it need to be useful on this task? What is the failure mode if it goes wrong, and who bears the cost? None of these are procedural facts. They are judgements that get better with use, reflection, and repeated contact with the tool’s actual behaviour.

That kind of learning cannot be delivered in a two-hour workshop. Staff can leave the workshop knowing how to open the interface and type a prompt, and remain completely unable to decide when to trust the result.

What follows

The practical implication is that “we ran AI training for staff” is not, on its own, evidence that an organisation has built AI literacy. The more useful question is whether the organisation has created the conditions in which mental-model shift happens — which means time, real work, reflection on where the tool helps and where it misleads, and senior people modelling sceptical, context-aware use.

The corollary is that attempts to accelerate AI adoption by scheduling more training usually do not work. What tends to work better is selecting a small number of high-value, well-bounded use cases (see The mid-tier AI adoption threshold) and letting literacy grow from sustained contact with those cases, rather than from general-purpose instruction.